Providing Fine-grained Network Metrics for Monitoring Applications using In-band Telemetry

Abstract

Network monitoring is fundamental for the correct and expected functioning of today’s large computer networks, as it allows network operators to identify disruptive flows, such as microbursts and elephant flows. In-band Network Telemetry (INT) has become one of the main tools for collecting network information in recent years. By piggybacking information using data plane packets, INT can deliver real-time network statistics to monitoring applications. However, INT’s fine granularity comes with a high network overhead cost, especially when monitoring high-throughput flows. Knowing this limitation, this paper focuses on accurately collecting network statistics using INT while keeping the telemetry overhead to a minimum for two monitoring applications: microburst and elephant flow detection. To this end, we present DINT, a Dynamic INT algorithm capable of collecting fine-grained network metrics with minimum telemetry overhead that adapts itself to the latest network developments. We evaluated DINT against two other algorithms for the microburst and the elephant flow monitoring scenarios. The evaluation results showed that DINT offers higher adaptability than other techniques, providing a more accurate network view while requiring fewer telemetry data and, consequently, improving the performance of the monitoring applications.